A scalable framework for synthesizing conditional porous media images using diffusion and GAN models

Generating micro-CT images of porous media is essential for rock characterization. However, such data are typically sparse and limited to small scanned intervals, leaving large sections without this valuable information. This paper presents a deep learning framework for generating large missing intervals of micro-CT data conditioned on petrophysical information. The workflow integrates the strengths of diffusion models for conditional image synthesis with generative adversarial networks for spatially consistent stitching of consecutive sub-images. The study investigates the performance of multiple conditional generative models and conditioning mechanisms to create micro-CT images that adhere closely to the petrophysical constraints. Additionally, several inpainting variants are evaluated to ensure high-quality stitching across sub-image boundaries. Model performance is assessed using four metrics including porosity, surface area, Euler number, and permeability. Furthermore, the large-scale micro-CT images are compared against real data using mean square differences between consecutive slices and their adherence to the conditional information. Overall, this study demonstrates the advantages of integrating diffusion and adversarial generative models to produce high-quality, conditionally consistent micro-CT images over large intervals from limited training data. The outcomes of this work have applications across multiple fields, including hydrogeology, geoscience, petroleum engineering, as well as hydrogen and carbon storage studies.